Project 01. The Engagement Charter

Nine months into Harbor Mutual's first attempt at a claims-correspondence assistant, a handler opened a finished letter and found it citing three policy clauses that did not exist in the customer's version of the wording. The archive held three versions of that clause with no flag for which one applied to the policy in front of you, and the tool had guessed by inception date and guessed wrong often enough that people noticed. Nobody filed an incident report. Handlers just stopped opening the drafts, quietly, over about two weeks. Three months later IT withdrew the integration, closing out a pilot nobody had formally killed.

Illustrative composite. Harbor Mutual and this incident are fictional, assembled for teaching; the same facts are narrated by Nowak in the discovery transcript below.

Harbor Mutual has agreed to a discovery workshop for the claims correspondence assistant. You ran it. The transcript below is what five stakeholders said in ninety minutes, and it contradicts itself in at least six places: two people claim to be the sponsor, the baseline is quoted three different ways, one person wants letters sent straight to customers and another says nothing leaves without a signature. Nobody in the room noticed. Your job is to notice, resolve each contradiction with a named owner, and hand over the deliverable. The first three files of your Architecture Dossier: the charter, the NFR budget sheet and the constraint register, built from a provided discovery transcript. Everything you need is on this page: the transcript, a stakeholder map and a data inventory. The deliverable is judged on whether the three files agree with each other and with what was actually said, not on polish.

Brief

Harbor Mutual is a regional property-and-casualty insurer under EU and UK supervision with an on-premises document archive and a cloud tenancy its analytics team set up. The chief operating officer wants an assistant that drafts claim correspondence grounded in policy wording and, she hopes, books adjuster visits. A first pilot ran for nine months and was cancelled — the failure above is that pilot. You are the architect. Use the templates from the three lessons in this module and the inputs below; do not invent facts the transcript does not support, and where a fact is missing write "cannot be filled yet" with the milestone that fills it.

Discovery workshop transcript

Illustrative composite. Harbor Mutual and every person, figure and system in this transcript are fictional and modeled for teaching. Attendees: Marta Ilves, chief operating officer; Daniel Okoro, head of claims; Priya Raman, chief information security officer; Tomasz Nowak, senior claims handler; Helen Marsh, finance business partner attending for the CFO. You are the architect.

Architect: Let's start with the problem, not the assistant. What changes if
this works?

Ilves: Cycle time. Our claims cycle time is the worst in our peer group and
the board has heard that twice. Every letter takes a handler twenty minutes,
we send thousands a week, and the assistant drafts them in seconds. I want the
drafts going straight out to customers so we get the full saving. That's the
pilot. And while we're at it, it should book the adjuster visits, because
that's the other thing handlers spend their day on.

Okoro: Nothing goes out to a customer without a handler's signature. That
isn't my preference, it's how we answered the regulator after the complaints
review last year. The handler reads it, the handler signs it.

Ilves: Then we lose half the saving.

Okoro: We lose half of a saving that doesn't exist yet. Booking is a different
problem anyway. The adjuster calendar is the bottleneck, not the handler.
Handlers wait days for a slot; a faster way to ask for one changes nothing.

Architect: Tomasz, the twenty minutes. Is that your number?

Nowak: No. Simple letters are templates, four minutes, and the case system
fills most of them. It's the complex ones: multi-peril, disputed cover,
anything where you have to find the right wording and cite it. Those are
eighteen minutes if the wording is where you expect, forty if it isn't. About
a third of what we send.

Architect: Why wouldn't the wording be where you expect?

Nowak: The archive has three versions of some wordings and no flag for which
one applies to the policy in front of you. You go by the inception date and
hope. That's what killed the last pilot, by the way. It cited clause numbers
that didn't exist in the customer's version. People stopped using it after a
fortnight. Nobody killed it; it just went quiet.

Ilves: IT killed it. They withdrew the integration.

Nowak: IT withdrew it three months after we'd all stopped.

Marsh: Can I ask about the money? The CFO wants a cost per letter and a saving
he can put in front of the board in eleven weeks. Budget approval is
quarterly, so if this isn't in the next cycle it waits three months. And I'll
be honest, the number he has in his head is headcount: fewer handlers.

Okoro: Then he can put it in front of the board without me. Headcount is
fixed; we're carrying a backlog of open claims, not surplus handlers. If this
is sold as a headcount cut the handlers won't touch it, and Tomasz will tell
you they didn't touch the last one for less.

Nowak: We didn't touch the last one because it lied about clauses. But yes,
nobody's going to train a tool that's meant to replace them.

Architect: Priya, constraints from your side.

Raman: Three. Inference and storage stay in the EU or the UK; no exceptions,
that comes from our regulatory posture, not from me. No vendor trains on our
data, and prompt retention stays inside whatever abuse-monitoring window is in
the contract, thirty days at the outside. And any new supplier goes through
third-party risk before we sign, which is ten weeks if the supplier is
responsive.

Ilves: We already have a cloud tenancy. The analytics team built it. Use that
and we skip the ten weeks.

Raman: That tenancy is in a US region. It's fine for anonymised analytics;
it's not fine for claim files. If the model provider is already an approved
sub-processor through that tenancy the assessment is shorter, but the region
has to change.

Architect: Who signs the letters legally, and who reviews the text?

Okoro: The handler signs. Legal reviews any template that goes to customers,
and they'd want to review anything generated that a customer reads. For
internal drafts they're not involved.

Architect: Identity. Does the assistant act as the handler or as a system
account?

Raman: As the handler. Every action in the case system is under the handler's
login through single sign-on. No service accounts touching claim files; that
was the finding from the last audit.

Architect: Who runs it after go-live?

Raman: The platform team. Three engineers. They run the case system and the
archive. None of them has run a model in production, so whatever this is, it's
a managed service.

Architect: Helen, what's the number that gets this into the next cycle?

Marsh: A cost per letter well under what a handler-minute costs, and a
measured saving. If the saving is a guess, it waits.

Architect: What would make each of you cancel at the pilot gate?

Okoro: A wrong citation reaching a customer. One.

Nowak: Handlers not using it by week four.

Raman: Any data outside the region.

Marsh: No measured saving by week eight.

Ilves: I don't plan to cancel it. Let's not write that down.

Architect: I'll write it down as an open question with your name on it.

Stakeholder map

The map below records who wants what and who can stop what. Read it before the charter, because the sponsor question is the first contradiction to resolve.

Data inventory

The Rule This Lab Tests

Use a stakeholder's number as a charter fact when it already carries a measurement method and a date; otherwise write "cannot be filled yet" with the milestone that fills it. Y is observable in your own transcript or discovery notes: does the number come with a "how it was measured" and a "when," or only a speaker's confidence?

Applying this to Harbor Mutual: Ilves's twenty-minute figure and Nowak's four/eighteen/forty split both fail the test — neither carries a measurement date — so both stay out of the baseline field even though Nowak's split is the more useful number for deciding which letters to automate.

Steps

Step 2's sponsor field is worked in full below; step 3's simple-letter templating row is faded — fill it yourself, then check the answer; steps 1, 4 and 5 stay open, resolved from your own reading of the transcript.

1. Read the transcript twice and list every contradiction. For each, record the two statements, the person who resolves it, and your resolution in one line. Expect at least six: the sponsor, the baseline, customer-facing output, the booking feature, the US tenancy, the headcount framing, and why the first pilot died. Keep this as discovery-notes.md in your working folder; it is not a Dossier file, but the acceptance checklist reads it.

2. Write dossier/00-charter.md from the Lesson 01 template. Name one sponsor and say why. State the problem in terms of complex letters, not correspondence. Put the baseline as a method and a date, not the twenty-minute figure. Write booking, customer-facing output and subrogation as non-goals with reasons drawn from the transcript. Write the exit evidence as the two measurements the room agreed to and the week they are taken. Put the vendor assessment in the milestones as a parallel track. Record Ilves's refusal to name a kill criterion as an open question with her name on it.

   Worked example — the sponsor field, filled in full from the transcript:

   markdowncopy

   ## Sponsor
   Marta Ilves, chief operating officer. She owns the target the pilot
   moves (cycle time) and she is the one who presents outcomes to the
   board — the two things a charter's sponsor field has to name. Helen
   Marsh's quarterly budget approval is a funding gate, not sponsorship;
   it is registered as a constraint owned by Marsh instead.
   

3. Write dossier/01-nfr-budgets.md from the Lesson 02 template. Score the four use cases the room raised: complex-letter drafting, booking, customer-facing drafts, simple-letter templating. Fill every budget row with a number, a measurement and an owner named in the stakeholder map; mark any number you set without an example as provisional. Add a row for each cancellation condition the room gave you, because each is a budget with a threshold of zero or a date.

   The complex-letter drafting row is worked below; fill the simple-letter templating row yourself before checking.

   Use case	Budget	Measurement	Owner
   Complex-letter drafting	Citation accuracy: no wrong or fabricated clause in more than 1 in 100 drafts	Quality team's fifty-letter quarterly audit, extended to the pilot's letters	Daniel Okoro

4. Write dossier/02-constraints.md from the Lesson 03 template. Profile the segment in four lines. Register at least ten constraints; every row's source must be a statement in the transcript, a cell in the data inventory, or an obligation you name. Register the US tenancy as soft with "cannot be used for this system." Register the archive's versioning problem as a hard data constraint with the archive owner as owner, even though the archive owner was not in the room.

5. Cross-check the three files. Every exit-evidence line in the charter has a budget row that says "from charter." Every non-goal has either a scorecard verdict or a register row. Every open question has an owner. Every owner in the budget sheet and the register appears in the stakeholder map or is marked as someone to invite.

Acceptance

• discovery-notes.md lists at least six contradictions, each with the resolving owner and a one-line resolution.
• The charter names one sponsor and gives the reason in one sentence.
• The charter's problem statement mentions complex letters and citation errors and contains no solution words.
• The baseline field states a method and a date; it does not use the twenty-minute figure as a fact.
• Non-goals include booking, customer-facing output and subrogation, each with a reason traceable to the transcript.
• Exit evidence names two measurements and the week they are taken, with thresholds deferred to the budget sheet.
• Milestones include the vendor risk assessment as a parallel track and the exit meeting's attendees.
• Ilves's refusal to name a kill criterion appears as an open question with her name.
• The scorecard scores four use cases against thresholds written above the table.
• Every budget row has a number, a measurement and an owner from the stakeholder map; provisional numbers are marked.
• The four cancellation conditions from the room appear as budget rows or kill criteria.
• The register has at least ten rows, each with a source from the transcript, the inventory or a named obligation, an owner, hard or soft, and what it fixes.
• The US tenancy and the archive versioning problem are registered.
• The cross-check in step five passes with no orphan exit evidence, non-goal, open question or owner.

Rubric

Key Takeaways

• Read a transcript for contradictions first; each one is a decision with an owner.
• Name one sponsor by who owns the target, not by who signs the cheque.
• Put the baseline in as a method and a date; a stakeholder's figure is a claim.
• Turn every cancellation condition the room gives you into a budget row or a kill criterion.
• Register constraints from absent owners and the data inventory, not just from the room.

What You Have Now

The first three files of your Architecture Dossier: the charter, the NFR budget sheet and the constraint register, built from a provided discovery transcript. They sit at dossier/00-charter.md, dossier/01-nfr-budgets.md and dossier/02-constraints.md, and together they close the Frame stage of the Frame → Decide → Prove → Run loop for Harbor Mutual. The next module reads all three: the model-selection record filters providers by the residency and retention rows, the grounding-data design starts from the archive's versioning constraint, and the agent specification inherits the rule that the assistant acts as the handler and takes no external action. If you built the module's files for your own case instead, compare them against these three and note which contradictions your own workshop never surfaced.